Project · Build a context budget calculator

Build a small tool that, given a model window, a system prompt, tool schemas, and a target output reservation, reports the variable token budget and a recommended history/retrieval split. Run it across three real features you care about and write up which lever (select/order/compress/cache/measure) each feature needs most.

• A `budget(window, system, tools, output_reservation)` function using the provider's real tokenizer.
• A short report applying it to 3 features, each labeled minimal/light/heavy context-engineering need.
• For each feature, the single highest-priority lever and why.

• Budget math uses a real tokenizer, not characters/4.
• Each feature is correctly classified by the dynamic/large/growing test.
• The recommended lever follows the priority order (select/order before model/prompt).